Super-huge section bearing steel dynamic recrystallization prediction method and system based on multi-modal data fusion and physical constraint deep learning

Abstract

The invention discloses an oversize section bearing steel dynamic recrystallization prediction method and system based on multi-modal data fusion and physical constraint deep learning, and the method comprises the steps: firstly constructing a multi-modal data fusion frame, and taking a dynamic recrystallization kinetics Avrami equation and a grain growth energy equation as constraint conditions to design a physical constraint deep neural network; a real-time control system is constructed by combining a high-temperature strain gauge array and a hydraulic servo system, so that the dynamic recrystallization rate, the grain size and the structure uniformity of the super-huge section bearing steel can be predicted, and the method is suitable for dynamic recrystallization behavior online monitoring and process regulation and control in the high-temperature deformation process of bearing steel hot rolling, forging and the like.

Description

Technical Field

[0001] The present invention relates to the field of intelligent technology for hot processing of metal materials, and specifically to a method and system for predicting dynamic recrystallization of extra-large cross-section bearing steel based on multimodal data fusion and physical constraint deep learning. Background Art

[0002] As high-end equipment such as wind turbine main shafts and nuclear power equipment develop towards larger sizes, accurate prediction of the dynamic recrystallization behavior of bearing steel with extra-large cross-sections has become a core requirement for improving material performance. The dynamic recrystallization process directly determines the degree of grain refinement and microstructure uniformity, which in turn affects the fatigue life of bearings. However, extra-large cross-section steels face significant non-uniform deformation problems during hot working. The cross-section temperature gradient is very high, resulting in significant differences in the recrystallization rate distribution, excessive grain size fluctuations, and low product qualification rates. The industry urgently needs a high-precision, real-time, low-cost dynamic recrystallization prediction technology to achieve online optimization and control of process parameters.

[0003] The prediction of dynamic recrystallization behavior in existing technologies has long relied on empirical models such as the Sellars equation and the Avrami kinetic equation. Such models are based on the assumption of uniform deformation and constant temperature, and are unable to characterize the non-uniform recrystallization behavior of bearing steel with extra-large cross-sections caused by temperature gradients and strain rate fluctuations during actual hot processing. Moreover, thermal simulation experiments are limited by the size of the specimen, resulting in high single-time costs. Finite element simulations require the division of millions of grid cells, and a single calculation takes 6-8 hours, which seriously lags behind the real-time control requirements of the production line. Existing data-driven models mostly use a single data source, ignoring the collaborative analysis of microstructure evolution and physical field data; pure data-driven methods lack the constraints of physical laws, resulting in prediction results that violate the thermodynamic constitutive relationship of the material. In addition, the excessive number of model parameters leads to high edge inference latency, which cannot meet real-time control requirements.

[0004] The above technical defects make it difficult for existing solutions to balance the accuracy, real-time performance and industrial implementation costs of dynamic recrystallization prediction: the traditional empirical model has an error of over 30% and cannot guide process optimization; the finite element simulation takes more than 6 hours per time and cannot match the real-time needs of the production line; due to the single data and lack of physical laws, the prediction results of intelligent technology are unreliable and the hardware deployment cost is high. Summary of the Invention

[0005] The purpose of the present invention is to provide a dynamic recrystallization prediction method and system for extra-large cross-section bearing steel based on multimodal data fusion and physical constraint deep learning, which can predict the dynamic recrystallization rate, grain size and structural uniformity of extra-large cross-section bearing steel, and is suitable for online monitoring and process control of dynamic recrystallization behavior in high-temperature deformation processes such as hot rolling and forging of bearing steel.

[0006] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0007] In a first aspect, the present invention discloses a method for predicting dynamic recrystallization of super-large cross-section bearing steel based on multimodal data fusion and physical constraint deep learning, the method comprising the following steps:

[0008] S1, using timestamp synchronization technology to align process parameters, microscopic images and physical field data to construct a multimodal dataset with temporal-spatial correlation;

[0009] S2: Standardize the process parameters to eliminate dimensional differences and extract the process parameter features; extract dynamic information including grain nucleation rate and growth rate from the microscopic images to construct a 1024-dimensional spatiotemporal feature; reduce the physical field data to 128 dimensions to retain the key distribution characteristics of the stress-strain field; calculate the entropy of each modal information based on the entropy weight method and assign weights to generate a 1086-dimensional fusion feature matrix;

[0010] S3, through the bidirectional GRU and graph convolutional network to jointly model the temporal-spatial features, embedding the Avrami dynamic equation and grain growth energy as constraints, constructing a physical constraint deep neural network, inputting a 1086-dimensional fusion feature matrix, and outputting the dynamic recrystallization rate, average grain size and microstructure uniformity index;

[0011] S4 uses a two-stage training method to train the GRU module using historical process data, load pre-trained weights, and jointly optimize multimodal fusion features and physical constraints to train and optimize the model;

[0012] S5 quantizes and compresses the physical constraint deep neural network and deploys it to the edge segment, interacting with the industrial control system and feeding back the adjustment results to the rolling production line in real time.

[0013] Step S1 further comprises:

[0014] Using a high-temperature strain gauge array to collect real-time process parameters and cross-sectional dimensions, including rolling temperature and strain rate, a three-dimensional thermomechanical coupling model of the extra-large cross-section is constructed based on DEFORM finite element software, outputting stress / strain field distribution data as physical field data. Furthermore, SEM video streams of the extra-large cross-section are collected as microscopic image data.

[0015] Timestamp synchronization technology is used to align process parameters, microscopic images, and physical field data to construct a multimodal dataset with temporal-spatial correlation.

[0016] Step S2 further comprises:

[0017] Normalize the process parameters including rolling temperature and strain rate to eliminate dimensional differences;

[0018] The images in the SEM video stream are input into the ResNet-50 model, and the feature vectors before the last fully connected layer are extracted. The extracted feature vectors are normalized to obtain the image feature vectors. The SEM video stream is then input into the 3D-CNN model frame by frame to extract the feature vectors of each frame. The feature vectors of each frame are average pooled to obtain the video feature vectors. The image feature vectors and video feature vectors are spliced ​​together to construct a 1024-dimensional spatiotemporal feature.

[0019] Use the t-SNE algorithm to reduce the dimensionality of the physical field data, retaining only the key distribution features of the stress-strain field and generating a 128-dimensional low-dimensional feature vector;

[0020] Based on the entropy weight method, the information entropy of each modal is calculated and weights are assigned. The normalized process parameters, 1024-dimensional spatiotemporal features and 128-dimensional low-dimensional feature vectors are fused according to the weights to generate a 1086-dimensional fusion feature matrix.

[0021] Furthermore, the expression of the 1086-dimensional fusion feature matrix is:

[0022] F fused =0.4F Process +0.3F micro +0.3F field

[0023] Where, F Process 、F micro and F field They are the normalized process parameter vector, 1024-dimensional spatiotemporal feature vector and 128-dimensional low-dimensional feature vector respectively.

[0024] Furthermore, the physical constraint deep neural network includes an input layer, a temporal processing unit, a spatial association unit and an output layer;

[0025] The input layer receives a 1086-dimensional fusion feature matrix; the time processing unit uses a 4-layer bidirectional GRU to capture the long-term and short-term dependencies of process parameters, and adopts a residual connection design to alleviate the gradient vanishing problem; the spatial association unit constructs a three-dimensional grid graph based on GCN, and aggregates local stress-strain field features through the graph attention mechanism; the output layer outputs the dynamic recrystallization rate, average grain size and microstructure uniformity index.

[0026] Furthermore, the total loss function of the physical constraint deep neural network is:

[0027] L Total =ω1L MSE +ω2L avrami +ω3L grain

[0028]

[0029] Among them, the Avrami dynamic equation L avrami Used to monitor the recrystallization rate change rate, grain growth Arrhenius equation L grain Used to constrain the grain size, ω1, ω2, and ω3 are weight factors corresponding to the minimum mean square error constraint, the recrystallization rate change constraint, and the grain size constraint, respectively; N represents the number of samples in the training batch, corresponding to multiple sets of hot working process parameters (such as strain rate and temperature) of the extra-large cross-section bearing steel specimen; X i represents the dynamic recrystallization volume fraction of the i-th sample (0≤X i ≤1), characterizing the degree of recrystallization of the material; t represents the time variable during the hot working process, which is related to the dynamic recrystallization kinetics evolution process of the bearing steel; k represents the dynamic recrystallization rate constant, which is related to the thermal activation process of the material microstructure evolution; d i represents the predicted grain size of the i-th sample (unit: μm), reflecting the grain refinement effect after recrystallization; d0 represents the initial grain size (unit: μm), corresponding to the original grain of the unrecrystallized bearing steel; Q represents the dynamic recrystallization activation energy (unit: J / mol), which is a material characteristic parameter of bearing steel; R represents the ideal gas constant (8.314 J / (mol\cdotpK)); Ti represents the absolute temperature of the i-th sample (unit: K), corresponding to the rolling or forging temperature in the hot working process.

[0030] In a second aspect, the present invention discloses a dynamic recrystallization prediction system for super-large cross-section bearing steel based on multimodal data fusion and physical constraint deep learning, the system comprising a multimodal data fusion module, a physical constraint deep network module and a real-time control module;

[0031] The multimodal data fusion module uses timestamp synchronization technology to align process parameters, microscopic images, and physical field data to construct a time-space-associated multimodal data set; standardizes process parameters to eliminate dimensional differences and extracts process parameter features; extracts dynamic information, including grain nucleation rate and growth rate, from microscopic images to construct 1024-dimensional spatiotemporal features; reduces the dimensionality of physical field data to 128 dimensions to retain key distribution features of the stress-strain field; calculates the entropy of each modal information based on the entropy weight method and assigns weights to generate a 1086-dimensional fusion feature matrix;

[0032] The physical constraint deep network module jointly models temporal-spatial features through a bidirectional GRU and a graph convolutional network, embeds the Avrami kinetic equation and grain growth energy as constraints, and processes the input 1086-dimensional fusion feature matrix to output dynamic recrystallization rate, average grain size, and microstructure uniformity index.

[0033] The real-time control module integrates edge computing equipment and industrial control systems to perform online prediction of the dynamic recrystallization state and closed-loop control of process parameters.

[0034] Furthermore, the real-time control module is composed of a data acquisition layer, an edge computing layer and a control execution layer to form a closed-loop system;

[0035] The data acquisition layer integrates a Phoenix infrared thermal imager and a Kistler piezoelectric sensor to capture process status in real time at a sampling rate of 1kHz. The edge computing layer uses NVIDIA Jetson AGX Xavier to deploy the PI-DNN model and achieves real-time prediction through the TensorRT acceleration engine. The control execution layer relies on the hydraulic servo system to dynamically adjust rolling parameters, and the control logic triggers process optimization based on the prediction results.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The dynamic recrystallization prediction method and system for extra-large cross-section bearing steel based on multimodal data fusion and physical constraint deep learning of the present invention is based on a multimodal data fusion and physical constraint deep learning framework. By embedding the Avrami kinetic equation and grain growth energy constraints, it eliminates purely data-driven non-physical solutions, covers an ultra-wide temperature range, and supports adaptive prediction of the size of extra-large cross-section bearing steel. It solves the core pain points of traditional methods such as large errors, slow speed, and high cost, and provides a new generation of technical paradigm for the quality control of hot processing of extra-large bearing steel. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flow chart of the dynamic recrystallization prediction method for extra-large cross-section bearing steel based on multimodal data fusion and physical constraint deep learning of the present invention. DETAILED DESCRIPTION

[0039] The embodiments of the present invention are described in further detail below with reference to the accompanying drawings.

[0040] The present invention discloses a method for predicting dynamic recrystallization of super-large cross-section bearing steel based on multimodal data fusion and physical constraint deep learning, the method comprising the following steps:

[0041] S1, using timestamp synchronization technology to align process parameters, microscopic images and physical field data to construct a multimodal dataset with temporal-spatial correlation;

[0042] S2: Standardize the process parameters to eliminate dimensional differences and extract the process parameter features; extract dynamic information including grain nucleation rate and growth rate from the microscopic images to construct a 1024-dimensional spatiotemporal feature; reduce the physical field data to 128 dimensions to retain the key distribution characteristics of the stress-strain field; calculate the entropy of each modal information based on the entropy weight method and assign weights to generate a 1086-dimensional fusion feature matrix;

[0043] S3, through the bidirectional GRU and graph convolutional network to jointly model the temporal-spatial features, embedding the Avrami dynamic equation and grain growth energy as constraints, constructing a physical constraint deep neural network, inputting a 1086-dimensional fusion feature matrix, and outputting the dynamic recrystallization rate, average grain size and microstructure uniformity index;

[0044] S4 uses a two-stage training method to train the GRU module using historical process data, load pre-trained weights, and jointly optimize multimodal fusion features and physical constraints to train and optimize the model;

[0045] S5 quantizes and compresses the physical constraint deep neural network and deploys it to the edge segment, interacting with the industrial control system and feeding back the adjustment results to the rolling production line in real time.

[0046] Specifically, in step 1, the rolling temperature, strain rate, and cross-sectional dimensions are collected in real time through a high-temperature strain gauge array, and a three-dimensional thermomechanical coupling model of an extra-large cross-section is constructed based on the DEFORM finite element software. The stress / strain field distribution data is output, and the timestamp synchronization technology is used to align the process parameters, microscopic images, and physical field data to construct a time-space associated multimodal data set.

[0047] In step 2, parameters such as temperature and strain rate are first normalized using Min-Max normalization to eliminate dimensional differences. Secondly, ResNet-50 and 3D-CNN are used to extract the spatiotemporal features of the SEM video stream (grain nucleation and growth temporal evolution). The t-SNE algorithm is then used to reduce the dimensionality of the finite element stress field data to generate a 128-dimensional low-dimensional feature vector. Finally, multi-source data are fused according to weight distribution (e.g., 40% for process parameters, 30% for microscopic features, and 30% for physical fields) to generate a 1086-dimensional fused feature matrix.

[0048] In step 3, the Avrami kinetic equation and the grain growth energy constraint conditions are embedded to construct a physical constraint deep neural network. The input layer has 1086-dimensional fusion features, and the hidden layer is set with 4 layers of gated recurrent units (GRU, 256 neurons) to process temporal dependencies + 2 layers of graph convolutional networks (GCN, 128 neurons) to model spatial associations. The output layer is set with dynamic recrystallization rate (0-1), average grain size (μm), and tissue uniformity index as the results.

[0049] In step 4, a two-stage training method is used, using historical process data and training GRU modules, loading pre-trained weights, and jointly optimizing multimodal fusion features and physical constraints. The learning rate is 1e -4 , batch size 64, set loss function, and train and optimize the model;

[0050] In step 5, the PI-DNN model is quantized, compressed, and deployed to the edge segment. It is linked to the PLC control system through the OPC UA protocol, and the adjustment results are fed back to the rolling production line in real time.

[0051] This paper also proposes a dynamic recrystallization prediction system for oversized bearing steel sections based on multimodal data fusion and physical constraint deep learning. By integrating process parameters, microstructure evolution data, and physical field simulation information in industrial scenarios, a multimodal fusion network and a physical constraint deep network (Physic-Informed DNN, PI-DNN) are constructed, and a real-time control module is linked to form a closed-loop optimization system. This method includes the following core modules, which act on the data acquisition network, feature extraction network, prediction network, and control network respectively:

[0052] Multimodal data fusion module (data acquisition network → feature extraction network): Responsible for integrating process parameters, microstructure evolution videos, and physical field data, achieving spatiotemporal alignment and dynamic weighted fusion to form a high-dimensional feature expression. Physical constraint deep network module (prediction network): Based on the bidirectional gated recurrent unit (GRU) and graph convolutional network (GCN), it constructs a time-space joint modeling network, embedding the Avrami kinetic equation and grain growth energy equation to ensure that the prediction results conform to the laws of material thermodynamics. Real-time control module (control network): Integrates edge computing devices and industrial control systems (PLC) to achieve online prediction of dynamic recrystallization states and closed-loop control of process parameters.

[0053] The multimodal data fusion module serves as the core of data preprocessing and feature extraction. It integrates process parameters, microstructure evolution sequences, and physical field simulation data, and constructs a high-dimensional feature space through spatiotemporal alignment and dynamic weighted fusion. First, the PTP protocol is used to achieve millisecond-level time synchronization of multi-source data (error ≤ 1ms), and the pixel coordinates of the SEM image are accurately matched with the finite element grid coordinates through the affine transformation matrix (spatial error ≤ 2μm). The process parameters are standardized by Z-Score to eliminate dimensional differences. The microstructure video is extracted through 3D-ResNet50 to extract 1024-dimensional spatiotemporal features (including dynamic information such as grain nucleation rate and growth rate). The physical field data is reduced to 128 dimensions by the t-SNE algorithm to retain the key distribution characteristics of the stress-strain field. In the dynamic weighted fusion stage, the entropy of each modal information is calculated based on the entropy weight method and weights are assigned (40% for process, 30% for micro, and 30% for physical field) to generate a 1086-dimensional fusion feature matrix, which is mathematically expressed as follows:

[0054] F fused =0.4F Process +0.3F micro +0.3F field

[0055] The multimodal data fusion module was verified in the rolling production line of a special steel enterprise. The prediction error was reduced by 50% under non-uniform deformation conditions, and the fusion time was ≤10ms, which is significantly better than the traditional single-modal method.

[0056] The Physical Constraint Deep Network (PI-DNN) module is the core architecture of the prediction network. It jointly models temporal and spatial features through a bidirectional GRU and a graph convolutional network (GCN), and embeds material thermodynamic constraint equations to ensure the physical rationality of the prediction results. The network input is a 1086-dimensional fused feature. The temporal processing unit uses a four-layer bidirectional GRU (256 neurons) to capture the long-term and short-term dependencies of process parameters. The residual connection design effectively alleviates the gradient vanishing problem. The introduction of residual connections prevents gradient vanishing and increases the network depth to four layers. The spatial association unit constructs a three-dimensional grid graph based on the GCN (128 neurons). The graph attention mechanism (GAT) aggregates local stress-strain field features. Edge weights are dynamically calculated based on the stress gradients of adjacent nodes:

[0057]

[0058] In terms of physical constraint mechanism, the Avrami kinetic equation (supervising the rate of change of recrystallization rate) and the grain growth Arrhenius equation (constraining grain size) are forcibly embedded in the loss function, and their expressions are:

[0059]

[0060] The total loss function is:

[0061] L Total =0.7L MSE +0.2L avrami +0.1L grain

[0062] In actual application, this module achieves a grain size prediction deviation of ±1.5μm (traditional method ±8μm) in the rolling process of 1200mm diameter bearing steel, and the model parameter volume is 15.8M.

[0063] The real-time control module, the execution end of process optimization, forms a closed-loop system consisting of a data acquisition layer, an edge computing layer, and a control execution layer. The data acquisition layer integrates a Phoenix infrared thermal imager (temperature measurement range 600-1300°C, ±1°C error) and a Kistler piezoelectric sensor (range 0-200 MPa, accuracy ±0.1% FS), capturing process status in real time at a 1kHz sampling rate. The edge computing layer uses the NVIDIA Jetson AGX Xavier to deploy the PI-DNN model, enabling real-time predictions through the TensorRT acceleration engine. The control execution layer leverages a Bosch Rexroth hydraulic servo system (response time ≤50ms) to dynamically adjust rolling parameters. The control logic triggers process optimization based on the prediction results.

[0064] In the exception handling mechanism, a Kalman filter suppresses sensor noise (SNR improvement ≥ 20dB), and an expert system reviews grain size anomalies (d < 20μm or d > 50μm). Application on a production line demonstrated that this module achieved temperature control accuracy of ±2°C and reduced strain rate fluctuation to ±0.1s-1, saving over 5 million yuan in annual manual intervention costs.

[0065] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0066] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A dynamic recrystallization prediction method for oversized cross-section bearing steel based on multimodal data fusion and physical constraint deep learning, characterized in that: The method The following steps are involved: S1, using timestamp synchronization technology to align process parameters, microscopic images and physical field data to construct a multimodal dataset with temporal-spatial correlation; S2, standardize the process parameters to eliminate dimensional differences and extract the process parameter characteristics; Extract dynamic information including grain nucleation rate and growth rate from microscopic images to construct 1024-dimensional spatiotemporal features; reduce the dimensionality of physical field data to 128 dimensions to retain key distribution features of the stress-strain field; Based on the entropy weight method, the information entropy of each modality is calculated and weights are assigned to generate a 1086-dimensional fusion feature matrix; S3, through the bidirectional GRU and graph convolutional network to jointly model the temporal-spatial features, embedding the Avrami dynamic equation and grain growth energy as constraints, constructing a physical constraint deep neural network, inputting a 1086-dimensional fusion feature matrix, and outputting the dynamic recrystallization rate, average grain size and microstructure uniformity index; S4 uses a two-stage training method to train the GRU module using historical process data, load pre-trained weights, and jointly optimize multimodal fusion features and physical constraints to train and optimize the model; S5 quantizes and compresses the physical constraint deep neural network and deploys it to the edge segment, interacting with the industrial control system and feeding back the adjustment results to the rolling production line in real time.

2. The method for predicting dynamic recrystallization of super-large cross-section bearing steel based on multimodal data fusion and physical constraint deep learning according to claim 1 is characterized in that: Step S1 further comprises: Using a high-temperature strain gauge array to collect real-time process parameters and cross-sectional dimensions, including rolling temperature and strain rate, a three-dimensional thermomechanical coupling model of the extra-large cross-section is constructed based on DEFORM finite element software, outputting stress / strain field distribution data as physical field data. Furthermore, SEM video streams of the extra-large cross-section are collected as microscopic image data. Timestamp synchronization technology is used to align process parameters, microscopic images, and physical field data to construct a multimodal dataset with temporal-spatial correlation.

3. The method for predicting dynamic recrystallization of super-large cross-section bearing steel based on multimodal data fusion and physical constraint deep learning according to claim 1 is characterized in that: Step S2 further comprises: Normalize the process parameters including rolling temperature and strain rate to eliminate dimensional differences; The images in the SEM video stream are input into the ResNet-50 model, and the feature vectors before the last fully connected layer are extracted. The extracted feature vectors are normalized to obtain the image feature vectors. The SEM video stream is then input into the 3D-CNN model frame by frame to extract the feature vectors of each frame. The feature vectors of each frame are average pooled to obtain the video feature vectors. The image feature vectors and video feature vectors are spliced ​​together to construct a 1024-dimensional spatiotemporal feature. Use the t-SNE algorithm to reduce the dimensionality of the physical field data, retaining only the key distribution features of the stress-strain field and generating a 128-dimensional low-dimensional feature vector; Based on the entropy weight method, the information entropy of each modal is calculated and weights are assigned. The normalized process parameters, 1024-dimensional spatiotemporal features and 128-dimensional low-dimensional feature vectors are fused according to the weights to generate a 1086-dimensional fusion feature matrix.

4. The method for predicting dynamic recrystallization of super-large cross-section bearing steel based on multimodal data fusion and physical constraint deep learning according to claim 2 is characterized in that: The expression of the 1086-dimensional fusion feature matrix is: F fused =0.4F Process +0.3F micro +0.3F field Where, F Process 、F micro and F field They are the normalized process parameter vector, 1024-dimensional spatiotemporal feature vector and 128-dimensional low-dimensional feature vector respectively.

5. The method for predicting dynamic recrystallization of super-large cross-section bearing steel based on multimodal data fusion and physical constraint deep learning according to claim 1 is characterized in that: The physical constraint deep neural network includes an input layer, a temporal processing unit, a spatial association unit and an output layer; The input layer receives a 1086-dimensional fusion feature matrix; the time processing unit uses a 4-layer bidirectional GRU to capture the long-term and short-term dependencies of process parameters, and adopts a residual connection design to alleviate the gradient vanishing problem; the spatial association unit constructs a three-dimensional grid graph based on GCN, and aggregates local stress-strain field features through the graph attention mechanism; the output layer outputs the dynamic recrystallization rate, average grain size and microstructure uniformity index.

6. The method for predicting dynamic recrystallization of super-large cross-section bearing steel based on multimodal data fusion and physical constraint deep learning according to claim 5, characterized in that: The total loss function of the physical constraint deep neural network is: L Total =ω1L MSE +ω2L avrami +ω3L grain Among them, the Avrami dynamic equation L avrami Used to monitor the recrystallization rate change rate, grain growth Arrhenius equation L grain Used to constrain the grain size, ω1, ω2, and ω3 are weight factors corresponding to the minimum mean square error constraint, the recrystallization rate change constraint, and the grain size constraint, respectively; N represents the number of samples in the training batch, corresponding to multiple sets of hot working process parameters for extra-large cross-section bearing steel specimens; X i represents the dynamic recrystallization volume fraction of the i-th sample, 0≤X i ≤1, characterizing the degree of recrystallization of the material; t represents the time variable during the hot working process, which is related to the dynamic recrystallization kinetics evolution process of the bearing steel; k represents the dynamic recrystallization rate constant, which is related to the thermal activation process of the material microstructure evolution; d i represents the predicted grain size of the i-th sample, reflecting the grain refinement effect after recrystallization; d0 represents the initial grain size, corresponding to the original grain of the unrecrystallized bearing steel; Q represents the dynamic recrystallization activation energy, which is a characteristic parameter of the bearing steel material; R represents the ideal gas constant; T i represents the absolute temperature of the i-th sample, corresponding to the rolling or forging temperature in the hot working process.

7. A dynamic recrystallization prediction system for large-section bearing steel based on multimodal data fusion and physical constraint deep learning, characterized in that: The system includes a multimodal data fusion module, a physical constraint deep network module and a real-time control module; The multimodal data fusion module uses timestamp synchronization technology to align process parameters, microscopic images, and physical field data to construct a multimodal dataset with temporal-spatial correlation; and standardizes the process parameters to eliminate dimensional differences and extract process parameter features; Dynamic information, including grain nucleation rate and growth rate, is extracted from microscopic images to construct a 1024-dimensional spatiotemporal feature. The physical field data is reduced to 128 dimensions to retain the key distribution characteristics of the stress-strain field. The entropy of each modal information is calculated and weighted based on the entropy weight method to generate a 1086-dimensional fusion feature matrix. The physical constraint deep network module jointly models temporal-spatial features through a bidirectional GRU and a graph convolutional network, embeds the Avrami kinetic equation and grain growth energy as constraints, and processes the input 1086-dimensional fusion feature matrix to output dynamic recrystallization rate, average grain size, and microstructure uniformity index. The real-time control module integrates edge computing equipment and industrial control systems to perform online prediction of the dynamic recrystallization state and closed-loop control of process parameters.

8. The dynamic recrystallization prediction system for super-large cross-section bearing steel based on multimodal data fusion and physical constraint deep learning according to claim 7 is characterized in that: The real-time control module consists of a data acquisition layer, an edge computing layer, and a control execution layer to form a closed-loop system; The data acquisition layer integrates a Phoenix infrared thermal imager and a Kistler piezoelectric sensor to capture process status in real time at a sampling rate of 1kHz. The edge computing layer uses NVIDIA Jetson AGX Xavier to deploy a physical constraint deep network module and achieves real-time prediction through the TensorRT acceleration engine. The control execution layer relies on the hydraulic servo system to dynamically adjust rolling parameters, and the control logic triggers process optimization based on the prediction results.

Images